improvement(container-runtime): For resubmit, don't submit ID Allocation op until submitting another op

[markfields]

Description

tl;dr - Delay submitting ID Allocation Ops until right before submitting other ops. Specifically updating the ID Allocation Op generated before replaying pending state.

Background

ID Allocation ops must be submitted before other ops to ensure compressed IDs contained in those ops can be understood by all clients. This change focuses on a special case before replaying pending state where the ID Compressor needs to re-take all unfinalized creation ranges, since any original ID Allocation ops (if any) won't be resubmitted.

Today, this op is submitted directly to the Outbox under a different codepath than we really expect to be submitting ops. In fact, we recently fixed a bug here where we don't schedule a flush, so it's possible if the user stops interacting with the document at this moment it will not be submitted at all.

The fix is to still generate the op before replaying pending states (so ID Compressor internal state is correct), but don't actually submit the op until we know we have reason to. (It's ok if we generate it, then drop it, and generate it again)

One other motivation for this change: ID Allocation ops are special because they are not "dirtyable" meaning they alone don't mark the container as dirty. We've found that the dirty/saved logic is a bit inconsistent though (see the struggle in #24646), so it's advantageous to avoid submitting these ops apart from other "dirtyable" ops. In fact we could choose follow up and stop marking them as "non-dirtyable" after this PR.

Prerequisite bug fix in Pending State Manager

There's a fix in PendingStateManager that became necessary after the main change here, it should have been this way all along. Since ID Allocation ops aren't resubmitted, we don't care to track their batch IDs. This means we shouldn't include batchId when replaying them (they're going to get dropped anyway) - without this fix we were triggering the empty batch logic (since they're not actually submitted on replay) when we shouldn't be resubmitting an empty batch.

[markfields]

Postponing. This was a precursor to #24510 but we went in a different direction with #24614

[markfields]

This is not the priority at the moment, but I do hope to merge this to move us towards consolidating logic around op submission/batching.

Maybe after the next release, there is some FUD about unintended side effects, i.e. whether the existing test coverage is sufficient since the behavior is so nuanced.

[markfields]

May revisit this in the fall. Closing for now.

Note, the last commit was a quick experiment. It clearly failed :)

[markfields]

Spiritual predecessor to #26784